Nemotron-Personas

Nemotron-Personas is a set of datasets of invented people, one per country, from NVIDIA.^1,3 Each record is an adult with structured attributes, such as age, sex, marital status, education, occupation and place, plus several short paragraphs describing them: their working life, sports, arts, travel, food, cultural background, skills, hobbies and goals.^1,2 Unlike a persona list written freely by a language model, the attributes are first drawn from a statistical model fitted to the country's census, so the ages, education levels and jobs come out in roughly the real proportions.^1,2 A language model then writes the text around them.^1,2 Minors are excluded.^1,2

The first, for the United States, appeared on 9 June 2025 and was built from the American Community Survey and a published dataset of names by race and ethnicity.¹ Version 1.1, on 28 October 2025, grew it to one million records holding about six million persona descriptions, spread over about 29,000 postcode areas.¹ Datasets for eleven more countries followed, the latest for Germany and the UK in October 2026.³ The UK one was built with Locai Labs from the 2021 and 2022 censuses of the four nations.² The US and UK sets were made with NeMo Data Designer, NVIDIA's open-source (Apache-2.0) synthetic data tool,^1,2,5 whose person sampler can draw from extended versions of these datasets for supported countries.⁴

The datasets are CC BY 4.0 and free for commercial use, and load in one line with the Hugging Face datasets library.^1,2 NVIDIA states the purpose as training and evaluating language models: more varied synthetic training data, less bias, and avoiding the decline in quality that comes from training models on other models' output.^1,2

The limits are in the cards. The statistical model behind the US data is proprietary and makes simplifying assumptions; for example, occupation is taken as independent of postcode once education, age and sex are known.¹ Sensitive fields such as ethnicity, religion and disability are used in the UK sampling but left out of the release, and names and addresses are left out of both.^1,2 The persona text is written by a language model and inherits its habits.^1,2 The cards measure how well the personas match census proportions, not whether they answer questions the way real residents would,^1,2 and they are not built as a complete population of any one town.